zjmletang
commented
7 months ago explanation about the netkvm patch.
- Problem Description see issue #972 and #942
-
Solution Although these are two separate issues, the solution is the same. In general, we will add a macro called ALICLOUD_MAX_FRAGMENTS_IN_ONE_NB to indicate the maximum number of fragments in one network packet. Its value depends on the specific vendor, so we define this macro in the Driver.AliCloud.props file. During the initialization stage of the driver, we will pre-allocate a shared memory. When the number of fragments in a sending packet exceeds the maximum, we will copy the sending packet to the shared memory using the copy method, and then transmit this block of memory to the network card.
- Implementation Details Q1 Where is this macro that defines the fragmentation limit defined? A: We have defined a file called "Driver.AliCloud.props" which is used to define AliCloud-specific information, so that the compiled driver includes AliCloud-specific details. Q2 Can users modify the maximum fragment size in the system, and how can it be modified? A: We can add the *MaxFragments value in the registry to modify the fragment count limit. By modifying this value, we can change the fragmentation limit. By default, this value is not present in the registry, and in such cases, the value of MaxFragments is set to the macro value ALICLOUD_MAX_FRAGMENTS_IN_ONE_NB. Why not do it in the driver parameters then? There are two considerations for this. Firstly, this parameter is usually related to the manufacturer and is not likely to change. Secondly, the meaning of this parameter seems more like an internal parameter and is not suitable for public disclosure.
Q3 How is the shared memory designed for the purpose of copying? A: For each network card device, we allocate 1024 (This value is worth discussing and considering.) physical pages, which are shared among all queues. This is done to minimize the impact on system performance caused by excessive shared memory allocation. In typical scenarios, 1024 physical pages are sufficient. In extreme cases where 1024 pages are not enough, we avoid using the copy mode and instead leverage the built-in TCP retransmission mechanism to mitigate the issue.
Q4 Why is the copy mode conditionally enabled only when NDIS_SUPPORT_NDIS650 is defined? A: This is because the probability of this issue occurring is low before the definition of NDSI_SG_DMA_V3_HAL_API (although it can still happen). Whether to add the restriction of NDIS_SUPPORT_NDIS650 is also worth discussing. I added this restriction to minimize the impact of changes as much as possible.
Q5 Why are some newly added variables in CNB not protected by locks? A: These variables appear to have a specific order of access, and it is unlikely for multiple threads to access them simultaneously. For performance reasons, we have chosen not to add locks.
YanVugenfirer
ybendito
kostyanf14
…fragments in a network packet
Signed-off-by: Zhang JianMing zhangjianming.zjm@alibaba-inc.com
for the problem description please visit #972 and #942